Back to home

Research Contributions

A deeper dive into the problems I care about (accessibility, security, and AI forensics) and the work I’ve done across papers, benchmarks, and a granted patent.

H-Index: 3•Total Citations: 21+•Google Scholar

Timeline

Federated learning schematic

Rethinking Data Integrity in Federated Learning: Are we ready?

2022
IEEE International WIE ConferenceFederated LearningSecurityPrivacy•
IEEE Xplore

S Dixit, PN Mahalle, GR Shinde

An investigation into the vulnerabilities of decentralized learning systems, specifically focusing on how malicious clients can poison global models.

Contributions

  • Conducted a comprehensive threat audit of current Federated Learning protocols.
  • Analyzed the trade-off between client privacy and the ability of the central server to verify data integrity.
  • Proposed a robust aggregation strategy to mitigate the impact of adversarial gradient updates.

Abstract

Investigates vulnerabilities in distributed learning—especially poisoning and data tampering—and proposes protocols to improve integrity in federated aggregation.

Hand reading Braille

Assistance Platform for Visually Impaired Person using Image Captioning

2023
Indian Patent OfficePatentComputer VisionAccessibility

Inventor: Shreyas Dixit

A patented wearable system that leverages real-time image captioning to provide a descriptive audio narrative of the user's environment.

Contributions

  • Conceptualized the hardware-software integration for low-latency scene description.
  • Implemented a lightweight CNN-LSTM architecture suitable for edge deployment on assistive devices.
  • Successfully navigated the patent filing process (No. 202321004399) focusing on the unique 'real-time narrative' feedback loop.

Abstract

The platform converts visual information into descriptive audio via image captioning, enabling users to understand surrounding scenes through hands-free narration.

Ocean waves

Wave-Former: Lag Removing Univariate Long Time Series Forecasting Transformer

2024
Ocean Engineering (Elsevier), Vol 312TransformersTime-SeriesOceanography•
Journal

D Shreyas, D Pradnya

A specialized Transformer architecture designed specifically for the maritime industry to solve the 'phase-lag' problem in wave height and frequency prediction.

Contributions

  • Developed a custom attention mechanism that prioritizes temporal alignment over standard point-wise accuracy.
  • Validated the model on global ocean buoy data, demonstrating a significant reduction in forecasting delay.
  • Optimized the Transformer for long-horizon univariate forecasting where seasonal trends are chaotic.

Abstract

Designs a Transformer architecture to reduce lag (phase shift) in long time series forecasting, improving usability for ocean wave prediction where timing alignment is critical.

Fact Check Explorer screenshot

DeFactify 4: Counter Turing Test (Text & Image) Overview and Datasets

2024
Workshop Proceedings (DeFactify 4)WorkshopDatasetAI Detection

R Roy, G Singh, A Aziz, S Bajpai, N Imanpour, S Biswas, K Wanaskar, S Dixit

A dual-track workshop project (Text and Image) focused on identifying the fingerprints of LLMs and Diffusion models in synthesized media.

Contributions

  • Led the data integrity verification for the 'Human vs. AI' text corpus.
  • Synthesized the 'Counter Turing Test' metrics to provide a unified score for detection difficulty.
  • Collaborated on the cross-modal analysis, comparing how detection difficulty varies between textual and visual AI artifacts.

Abstract

A comprehensive overview and dataset release establishing benchmarks for human vs. AI-generated content detection via the DeFactify 4 workshop series.

Example photo with a visible watermark

PECCAVI: Visual Paraphrase Attack Safe and Distortion Free Image Watermarking

2025
CVPR 2026GenAIForensicsSecurity•
arXiv

S Dixit, A Aziz, S Bajpai, V Sharma, A Chadha, V Jain, A Das

A watermarking technique for AI-generated images that stays detectable under 'visual paraphrase' attacks—subtle, semantics-preserving edits that typically break standard watermarks.

Contributions

  • Designed a distortion-free embedding mechanism that maintains 100% visual fidelity while embedding robust metadata.
  • Modeled the 'Visual Paraphrase' threat vector, simulating real-world attacker behavior like slight cropping and color shifting.
  • Engineered the extraction algorithm to achieve high precision even after lossy compression.

Abstract

Proposes a watermarking approach that is robust against 'visual paraphrase' attacks while maintaining zero visual distortion in the source media, critical for verifying AI-generated content origins.

Presentation slide asking whether content is real or fake

The Visual Counter Turing Test (VCT²): A Benchmark for AI-Generated Image Detection

2025
Proceedings of the 14th IJCNLPBenchmarkingNLPMulti-modal

N Imanpour, A Borah, S Bajpai, S Ghosh, SR Sankepally, HM Abdullah, S Dixit

Introducing the Visual AI Index (V_AI), this work establishes a rigorous testing framework to evaluate if current detection systems can outpace rapidly evolving generative models.

Contributions

  • Co-developed the Visual AI Index (V_AI) metric to quantify the 'believability' gap in AI imagery.
  • Identified critical failure points in multi-modal models when faced with semantic inconsistencies.
  • Standardized the evaluation protocol for the IJCNLP community to ensure cross-model comparability.

Abstract

Introduces the Visual AI Index (V_AI) and a benchmark for evaluating AI-generated image detection, highlighting gaps in current detection pipelines against increasingly capable generative models.

Conceptual diagram of AI generation

A Comprehensive Dataset for Human vs. AI Generated Image Detection

2026
arXiv PreprintGenAIComputer VisionDataset•
arXiv

R Roy, N Imanpour, A Aziz, S Bajpai, G Singh, S Biswas, K Wanaskar, S Dixit

A massive-scale longitudinal study and dataset release addressing the difficulty of distinguishing hyper-realistic Diffusion and GAN-based imagery from authentic photography.

Contributions

  • Curated and validated thousands of high-fidelity image pairs across diverse domains.
  • Developed metadata schemas to track generator provenance and prompt engineering styles.
  • Established a standard evaluation pipeline for baseline detection models on the 2026 dataset release.

Abstract

Presents a state-of-the-art benchmark dataset for the 2026 landscape of generative AI, focusing on edge cases where human perception and automated detectors frequently fail.